Theorem 1, p. 385 — regret-balancing grid forecast is ε-calibrated against adaptive Nature
ProvedAsympCalib.Main.theorem_1Fix and an integer grid parameter with . Let a randomized forecaster choose a grid probability after each finite history according to a distribution satisfying the regret-flow conservation equations (2), computed from earlier rounds. Nature may use any randomized rule depending on past realized forecasts and outcomes, but does not observe the current forecast draw. If is the realized quadratic calibration score, then there is a time such that
This identifies the particular regret-balancing algorithm as an -calibrated forecaster against every adaptive Nature. It is stronger in content than merely asserting that some calibrated forecaster exists.
Formalization Note The paper writes ; the statement uses . Its limit condition is expressed as an eventual strict probability bound; the proof yields convergence of this probability to . The extra explicit prevents a degenerate grid. The grid has points, encoded as vectors , and the vectors are distinct. Round uses the forecaster distribution from the history prefix before , with represented by regret over that prefix. The event concerns the realized score , not .
import Mathlib import Definitions.Def_AsympCalib_Main_Setting noncomputable section
namespace AsympCalib.Main
theorem theorem_1 (ε : ℝ) (hε : 0 < ε)
(k : ℕ) (hk0 : 0 < k) (hk : 1 / ε ^ 2 ≤ (k : ℝ))
(σ : History → PMF (Fin (k + 1)))
(hσ : IsConservationForecaster k σ)
(A : History → PMF (Fin 2)) :
∃ t₁ : ℕ, ∀ t ≥ t₁,
ENNReal.ofReal (1 - ε) <
(CalibratedCE.Forecast.histLaw (forecaster k σ) A t).toOuterMeasure
{h : History | CalibratedCE.Forecast.calibScore2Hj h 1 < ε} := by sorry
end AsympCalib.Main
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.